Font Size: 

The rising prevalence of autism spectrum disorder (ASD) around the world affects an estimated 1 in 59 children. In Bangladesh, the prevalence is estimated to be similar, though weak healthcare infrastructure and lack of access to mental health resources makes it difficult to identify children at risk before age 4, when behavioral interventions are most beneficial. There is a huge disparity in ASD prevalence in Bangladesh between rural, urban low income, and urban middle-to-high income families (0.68 per 1000, 1-2 per 1000, and 30 per 1000 children, respectively). A mobilized autism risk initiative (AMARI) integrates machine-learning technology to screen for ASD in Bangladeshi children in a span of minutes outside of the clinical setting, therefore leap-frogging long waitlists and a lack of access to resources, by using two complimentary machine-learning classifiers: (1) a short parent-directed questionnaire, and (2) top-ranked feature extraction from short home video clips. Stanford University, in collaboration with Dhaka Shishu Children’s Hospital, collected home videos on 150 children (n=50 with ASD, n=50 with speech-language delay, and n=50 neurotypical) and tested the classification system’s performance. We confirmed that video raters and our machine learning models are able to quickly and accurately classify children with ASD from children without ASD across cultural and language barriers. The ability to dynamically screen for ASD in a diverse population while maintaining clinically optimal levels of accuracy, sensitivity, and specificity, across language and cultural barriers, provides a model to expand towards other conditions and paves a way towards improving healthcare globally.

autism; screening; mobile; diagnosis; children